acceptodds
Withdrawn from ICLR 2027

Single-Source Generalization for Cross-Geospatial Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-geospatial shifts in remote-sensing imagery cause large drops in detector performance. These shifts are compound, as environmental conditions and scene characteristics vary together across regions. Single-source domain generalization for object detection (SDGOD) has been extensively studied under weather and illumination shifts but remains under-examined for such compound shifts. To bridge this gap, we propose CoDAlign, which combines factor-guided composite augmentation with foreground-decoupled classification alignment. First, we organize source-image augmentation around explicit geographic appearance factors and validate the induced changes in an image-anchored factor space. The same factor axes enable controlled diagnostics of the robustness gained through augmentation and the shifts that remain challenging. Second, we identify how low foreground confidence in augmented images weakens class consistency in full-distribution prediction alignment. Our decoupled alignment strengthens correct-class consistency at regions matched to ground-truth objects while retaining foreground-background and non-target class alignment. Experiments across cross-region benchmarks and multiple detector architectures demonstrate that CoDAlign consistently outperforms existing SDGOD methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Domain Generalization, Object Detection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_maJbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a method for single-source domain-generalizable object detection in remote sensing imagery. The method aligns the fine-tuned object detector's predictions along original and augmented views of source-domain images. The augmented views are generated using prompt-driven image edits and traditional augmentation approaches. The approach is evaluated on two remote sensing object detection benchmark datasets that measure detection performance under distribution shifts and here yields strong performance compared to prior ResNet-50 Faster R-CNN-based methods.

Critical strengths

The evaluation compares to a wide range of relevant prior work. Factor edits are separately evaluated as additional augmentation method within prior methods. The ablation study clearly quantifies the contribution of each methodological component and includes confidence intervals. The factor-space diagnostics is interesting and highlights the value of foreground-decoupled classification alignment well.

Critical weaknesses

===Detectors=== The main experiments are performed with a ResNet-50 Faster R-CNN, an arguably outdated detector. While this might enable easy comparisons to prior works, this leaves open whether foreground-decoupled classification alignment is still effective on top of stronger, more modern detectors. Particularly informative would be a comparison using a DINOv3 backbone, which might enable direct comparisons of factor edits + decoupled knowledge distillation with Bridge.

===Benchmark Diversity=== The paper focuses on a very narrow set of benchmarks within domain-generalizable object detection, but it seems like the method would transfer just as well to other DGOD tasks. Since the paper primarily adapts existing ideas and methods such as DivAlign and Decoupled knowledge distillation, showing generalization to additional domains and tasks could strengthen the claim that this particular combination of methods is particularly useful.

===Compound shifts=== The abstract and introduction discuss compound shifts as a limitation of existing SDGOD approaches, but it is not explicitly discussed how this limitation is addressed here. The factor edits seem to be applied one at a time rather than in a compound fashion.

Questions for authors
  • Can you clarify if and how factor edits compound?
1 comment

Thank you for the review and the positive comments on our comparisons and diagnostics. Your suggestion to compare with Bridge on DINOv3 led to the new same-backbone experiments.

Detectors. We trained the full method on the DINOv3 ViT-L detector of Bridge and on a Swin-T Faster R-CNN. In the two settings of the submission, CoDAlign with DINOv3 reaches 12.17 on CZ and 25.95 on HE, against 12.03 and 21.35 for Bridge. Damaged-building AP$_{50}$ rises from 2.38 to 9.52 (general response). On Swin-T, DCA adds 2.05 points on HE over the same alignment without decoupling, with a gain in all three seeds.

Benchmark diversity. Each remote-sensing benchmark now has a second source region. For other DGOD tasks, we add DCA to two existing class-consistency losses, with PASCAL VOC as the source and Clipart, Comic, and Watercolor as targets. It adds 0.83 points of mean AP$_{50}$ to the KL-divergence loss and 0.45 to the Jensen-Shannon loss.

Relation to DivAlign and DKD. As one of the popular alignment methods, DivAlign supplies the two-view KL loss that we analyze, and DKD supplies the way of splitting a KL divergence. The finding is specific to two-view alignment in detection: the background class of the detector makes the class term of the alignment loss scale with foreground confidence, and augmentation lowers this confidence on labeled objects. DKD decouples the non-target-class term of a distillation loss from the teacher's confidence in the target class. DCA decouples the labeled-class term from foreground confidence, and only at regions matched to ground-truth objects.

Compound shifts. Each training image is edited along one factor at a time. The images in a batch are edited along different factors. We agree that the submission did not show how this addresses compound shifts as a distinct limitation. A future submission will make a more specific claim that the evidence supports: the edits cover named appearance factors along which the target regions differ from the source, and the diagnostics report robustness along each factor.

Reviewer_q2CLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a systematic study geared towards improving single-source domain generalization for object detection (SDGOD) under cross-geospatial appearance shifts. They propose CoDAlign which uses a remote sensing image with a foreground class and generates a corrupted image based on text prompts specifically made to change the environmental style while asking to maintain the original camera angle, object placement, geometric structure. They show that by using paired two-view training, the foreground class can be extracted from domain shifted data across multiple detectors over two cross-region benchmarks.

Critical strengths
  • The CoDAlign algorithm is easily understandable and can be generalized to different geospatial imagery
  • The benchmark and ablation results are comprehensive for a conference publication
Critical weaknesses
  • The authors claim they get state-of-the-art performance but their results show Cauvis and Bridge performing better than CoDAlign. I am not sure how they can claim as such
  • The use of AI editor for changing an image based on text prompts is the weak point of the method and should be acknowledged by the authors - specifically their emperical validation and quality checks are not sufficient to be considered as generalizable
  • In Table 2, it would be useful to check the overall image distribution in terms of KL divergence than the AP50 - they use KL divergence as measure in Eq 3 but have not provided the results
  • The proposed algorithm is simple enough to be an engineering application rather than a scientific breakthrough
Questions for authors
  • Clarify how the given algorithm is better than similar proposed algorithms?
  • How can the AI editor consistency be validated?
1 comment

Thank you for the review, and for finding the method easy to follow and the experiments comprehensive.

W1, state-of-the-art claim. The submission compared a ResNet-50 detector with ViT-L methods, and we agree that the claim was too broad for that comparison. A future submission will compare methods within each backbone. In the same settings as the submission, CoDAlign on the DINOv3 detector of Bridge reaches 12.17 on CZ and 25.95 on HE, against 12.03 and 21.35 for Bridge (general response).

W2 and Q2, the image editor. We agree that the method depends on an image editor, as many downstream methods depend on a pretrained foundation model, and a future submission will state this explicitly. Nothing in the pipeline is specific to one editor: the editor is frozen, is used only through text prompts, and can be replaced. The submission checks each factor before training. Its low, neutral, and high edits must be ordered along the factor's direction in an image-embedding space, and an edit must move the other factors less than its own. We also inspect object regions for destructive edits. Generic style edits from the same editor, under the same geometry-preservation instruction, reach 13.97 on HE where factor edits reach 16.60, so the editor alone does not explain the gain. We also tested a second editor with the same prompts. It moves the appearance around objects in the same directions, and detectors trained on the first editor's edits keep more of their accuracy on the second editor’s held-out edits. Newer editors can be substituted without changing the method. The alignment module does not depend on the editor and applies to any augmented view.

W3, KL divergence. Eq. 3 is a training loss: the KL divergence between the class predictions that the detector makes for the same region in the original and augmented views. This loss does not measure a distance between image distributions, so it gives no value to report in Table 2, which compares detectors by AP$_{50}$. Future versions of the paper will state this where the equation is introduced. The question of how edited images relate to real target images is addressed by the factor-space diagnostics, which place both on the same factor axes.

W4 and Q1, simplicity and comparison with similar algorithms. DCA changes only one specific weight in existing losses, and we regard this as a algorithmic contribution with no cost and decent generalization performance. We have since compared DCA directly with similar algorithms. Five simpler loss changes add at most 0.81 points on HE where DCA adds 1.22. Adding DCA to two existing class-consistency losses improves both on three artistic targets (general response).

Reviewer_pqWtOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies single-source domain generalization for object detection under compound cross-geospatial appearance shifts. CoDAlign (i) edits source images along four named appearance factors (vegetation, soil, haze, shadow) with a frozen text-guided editor, and (ii) combines factor-guided composite augmentation with foreground-decoupled classification alignment to diversify training appearances and strengthen class consistency. Extensive experiments across cross-region benchmarks demonstrate that CoDAlign achieves state-of-the-art performance.

Critical strengths
  1. The paper addresses a crucial and highly practical challenge in the community: improving the generalization capability of object detection models.
  2. The proposed method achieves leading quantitative metrics among reported results on the evaluated benchmarks.
Critical weaknesses
  1. The bibliography presentation is inconsistent throughout the paper. Even for CVPR references, the formatting styles vary significantly. Please homogenize the citation format.
  2. Most importantly, the proposed framework is essentially a data augmentation pipeline and appears to be largely an engineering contribution, lacking substantial algorithmic or technical innovation.
  3. On the xView: CZ dataset, the proposed data augmentation strategy brings very limited performance improvement, questioning its effectiveness under certain conditions.
  4. In Table 1, combining the top and bottom sections into a single table is misleading and creates unnecessary confusion for the reader. I suggest splitting or reorganizing it.
  5. The empirical evaluation is evaluated on only two datasets, which is insufficient to fully validate the broad effectiveness and generalizability of the proposed method.
Questions for authors
  1. Please address all the points and concerns raised in the Weaknesses section above.
1 comment

Thank you for the review, and for recognizing the practical importance of the problem.

W1, bibliography. We agree, and we will unify venue names and entry formats across the reference list in a future submission.

W2, engineering contribution. Please see the general response. The alignment module rests on an analysis of the standard two-view KL loss. The resulting change outperforms five simpler alternatives and also improves two other consistency losses on a benchmark outside remote sensing. We test the augmentation module as a separate component as well. Generic style edits from the same editor reach 13.97 on HE where factor edits reach 16.60, and adding factor edits to existing methods improves every host on both benchmarks (Table 2).

W3, limited gains on CZ. Gains on CZ are small for every method: all ResNet-50 rows of Table 1 lie within three points of each other. CoDAlign improves over the baseline of the same backbone on CZ with each backbone we tested. With the tropical source, the gains are 1.18 points over Faster R-CNN with ResNet-50, 0.79 with Swin-T, and 0.14 over Bridge with DINOv3. On the temperate source, which we added after submission, they are 1.58 with Swin-T and 0.18 over Bridge. Factor augmentation alone also improves every host in Table 2 on CZ.

W4, Table 1. We agree. In a future submission, Table 1 will group rows by backbone and place a CoDAlign row in each group, so that every comparison is made within one backbone.

W5, two datasets. The evaluation now covers six regions across the two benchmarks and four source domains: each benchmark has a second source region, with its other two regions as targets. The method is also evaluated with three backbones. DCA is also evaluated outside remote sensing, on PASCAL VOC to Clipart, Comic, and Watercolor (general response).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the three reviewers for their time and for the concrete suggestions. Since submission, we have added an analysis of why existing augmentation falls short under geographic shift, and its findings change how the paper is framed. The resulting changes are larger than a discussion-period update should carry, so we have decided to withdraw the submission. Before withdrawing, we record here how each concern has been handled, so that the discussion is complete for later readers. Replies to the individual reviews follow in separate comments.

What the paper contributes. CoDAlign has two parts. The first edits each source image along four named appearance factors (vegetation, soil, haze, shadow) with a frozen text-guided editor and reuses the boxes and labels of the source image. The second, foreground-decoupled classification alignment (DCA), follows from an analysis of the two-view KL alignment loss used in prior works. Let qaq_a be the detector's confidence, on the augmented view, that a region contains an object. The KL loss separates into a foreground term and a class term, and the class term is multiplied by qaq_a. Under this loss, nearly half of the labeled objects in held-out factor edits receive qa<0.2q_a < 0.2, and these objects show the largest class disagreement between the two views. The loss therefore aligns the class least where the views disagree on it most. DCA removes the weight qaq_a from the labeled-class term at regions matched to ground-truth objects. It is an effective algorithm that changes one weight and adds no parameters.

New analysis. A future submission will open with a diagnosis of the problem, from which the two modules follow. We measure the appearance around each labeled object along vegetation, soil, and local darkness. In the target regions, the surroundings of many labeled objects lie outside the range that the source region covers. On this cross-geospatial shift, the image-level augmentations of existing SDGOD methods seldom move an object's surroundings beyond that range, whereas factor edits do. This is also consistent with the gains that factor edits add on top of those methods (Table 2).

Same-backbone comparisons. Reviewers q2CL and maJb noted that the submission compared a ResNet-50 detector with ViT-L methods. We have since trained the full method on a Swin-T Faster R-CNN and on the DINOv3 ViT-L detector of Bridge. The table reports mean target-domain AP50 in the two settings of the submission; Dmg. is damaged-building AP50.

BackboneMethodCZ (source: Tropical)HE (source: Michael)HE Dmg.
ResNet-50Faster R-CNN9.3313.043.40
ResNet-50CoDAlign10.5122.739.02
Swin-TFaster R-CNN10.8616.833.73
Swin-TCoDAlign11.6526.5410.43
DINOv3 ViT-LBridge12.0321.352.38
DINOv3 ViT-LCoDAlign12.1725.959.52

The updated recipe drops the box-alignment term of the submission, so the ResNet-50 values differ from the submitted Table 1.

Evidence for DCA. On HE with the Michael source, DCA adds 1.22 points over the same alignment without decoupling and raises damaged-building AP50 from 6.92 to 9.02. Five simpler alternatives (reverse KL, JS divergence, extra cross-entropy on augmented objects, class-balanced cross-entropy, and focal loss) add at most 0.81 points and reach at most 7.42 on damaged buildings. DCA also improves other consistency losses. With PASCAL VOC as the source and Clipart, Comic, and Watercolor as targets, it adds 0.83 points of mean AP50 to a KL-divergence class-consistency loss and 0.45 to a Jensen-Shannon class-consistency loss, the two forms used by existing SDGOD methods.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.